Cursor has released a technical paper [1] and several blog posts [2] describing the continued pretraining and RL they do on top of Kimi K2.5.

It is true that they were not transparent about the base model that they used until the model slug was discovered by a Twitter user via the API.

[1]: https://arxiv.org/abs/2603.24477 [2]: https://cursor.com/blog/real-time-rl-for-composer

Kimi is the base, but they've done tons of finetuning on top to produce a really good completions model.